1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/content/media/webaudio/blink/ReverbConvolver.cpp Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,269 @@
1.4 +/*
1.5 + * Copyright (C) 2010 Google Inc. All rights reserved.
1.6 + *
1.7 + * Redistribution and use in source and binary forms, with or without
1.8 + * modification, are permitted provided that the following conditions
1.9 + * are met:
1.10 + *
1.11 + * 1. Redistributions of source code must retain the above copyright
1.12 + * notice, this list of conditions and the following disclaimer.
1.13 + * 2. Redistributions in binary form must reproduce the above copyright
1.14 + * notice, this list of conditions and the following disclaimer in the
1.15 + * documentation and/or other materials provided with the distribution.
1.16 + * 3. Neither the name of Apple Computer, Inc. ("Apple") nor the names of
1.17 + * its contributors may be used to endorse or promote products derived
1.18 + * from this software without specific prior written permission.
1.19 + *
1.20 + * THIS SOFTWARE IS PROVIDED BY APPLE AND ITS CONTRIBUTORS "AS IS" AND ANY
1.21 + * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
1.22 + * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
1.23 + * DISCLAIMED. IN NO EVENT SHALL APPLE OR ITS CONTRIBUTORS BE LIABLE FOR ANY
1.24 + * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
1.25 + * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
1.26 + * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
1.27 + * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
1.28 + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
1.29 + * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1.30 + */
1.31 +
1.32 +#include "ReverbConvolver.h"
1.33 +#include "ReverbConvolverStage.h"
1.34 +
1.35 +using namespace mozilla;
1.36 +
1.37 +template<>
1.38 +struct RunnableMethodTraits<WebCore::ReverbConvolver>
1.39 +{
1.40 + static void RetainCallee(WebCore::ReverbConvolver* obj) {}
1.41 + static void ReleaseCallee(WebCore::ReverbConvolver* obj) {}
1.42 +};
1.43 +
1.44 +namespace WebCore {
1.45 +
1.46 +const int InputBufferSize = 8 * 16384;
1.47 +
1.48 +// We only process the leading portion of the impulse response in the real-time thread. We don't exceed this length.
1.49 +// It turns out then, that the background thread has about 278msec of scheduling slop.
1.50 +// Empirically, this has been found to be a good compromise between giving enough time for scheduling slop,
1.51 +// while still minimizing the amount of processing done in the primary (high-priority) thread.
1.52 +// This was found to be a good value on Mac OS X, and may work well on other platforms as well, assuming
1.53 +// the very rough scheduling latencies are similar on these time-scales. Of course, this code may need to be
1.54 +// tuned for individual platforms if this assumption is found to be incorrect.
1.55 +const size_t RealtimeFrameLimit = 8192 + 4096; // ~278msec @ 44.1KHz
1.56 +
1.57 +const size_t MinFFTSize = 128;
1.58 +const size_t MaxRealtimeFFTSize = 2048;
1.59 +
1.60 +ReverbConvolver::ReverbConvolver(const float* impulseResponseData, size_t impulseResponseLength, size_t renderSliceSize, size_t maxFFTSize, size_t convolverRenderPhase, bool useBackgroundThreads)
1.61 + : m_impulseResponseLength(impulseResponseLength)
1.62 + , m_accumulationBuffer(impulseResponseLength + renderSliceSize)
1.63 + , m_inputBuffer(InputBufferSize)
1.64 + , m_minFFTSize(MinFFTSize) // First stage will have this size - successive stages will double in size each time
1.65 + , m_maxFFTSize(maxFFTSize) // until we hit m_maxFFTSize
1.66 + , m_backgroundThread("ConvolverWorker")
1.67 + , m_backgroundThreadCondition(&m_backgroundThreadLock)
1.68 + , m_useBackgroundThreads(useBackgroundThreads)
1.69 + , m_wantsToExit(false)
1.70 + , m_moreInputBuffered(false)
1.71 +{
1.72 + // If we are using background threads then don't exceed this FFT size for the
1.73 + // stages which run in the real-time thread. This avoids having only one or two
1.74 + // large stages (size 16384 or so) at the end which take a lot of time every several
1.75 + // processing slices. This way we amortize the cost over more processing slices.
1.76 + m_maxRealtimeFFTSize = MaxRealtimeFFTSize;
1.77 +
1.78 + // For the moment, a good way to know if we have real-time constraint is to check if we're using background threads.
1.79 + // Otherwise, assume we're being run from a command-line tool.
1.80 + bool hasRealtimeConstraint = useBackgroundThreads;
1.81 +
1.82 + const float* response = impulseResponseData;
1.83 + size_t totalResponseLength = impulseResponseLength;
1.84 +
1.85 + // The total latency is zero because the direct-convolution is used in the leading portion.
1.86 + size_t reverbTotalLatency = 0;
1.87 +
1.88 + size_t stageOffset = 0;
1.89 + int i = 0;
1.90 + size_t fftSize = m_minFFTSize;
1.91 + while (stageOffset < totalResponseLength) {
1.92 + size_t stageSize = fftSize / 2;
1.93 +
1.94 + // For the last stage, it's possible that stageOffset is such that we're straddling the end
1.95 + // of the impulse response buffer (if we use stageSize), so reduce the last stage's length...
1.96 + if (stageSize + stageOffset > totalResponseLength)
1.97 + stageSize = totalResponseLength - stageOffset;
1.98 +
1.99 + // This "staggers" the time when each FFT happens so they don't all happen at the same time
1.100 + int renderPhase = convolverRenderPhase + i * renderSliceSize;
1.101 +
1.102 + bool useDirectConvolver = !stageOffset;
1.103 +
1.104 + nsAutoPtr<ReverbConvolverStage> stage(new ReverbConvolverStage(response, totalResponseLength, reverbTotalLatency, stageOffset, stageSize, fftSize, renderPhase, renderSliceSize, &m_accumulationBuffer, useDirectConvolver));
1.105 +
1.106 + bool isBackgroundStage = false;
1.107 +
1.108 + if (this->useBackgroundThreads() && stageOffset > RealtimeFrameLimit) {
1.109 + m_backgroundStages.AppendElement(stage.forget());
1.110 + isBackgroundStage = true;
1.111 + } else
1.112 + m_stages.AppendElement(stage.forget());
1.113 +
1.114 + stageOffset += stageSize;
1.115 + ++i;
1.116 +
1.117 + if (!useDirectConvolver) {
1.118 + // Figure out next FFT size
1.119 + fftSize *= 2;
1.120 + }
1.121 +
1.122 + if (hasRealtimeConstraint && !isBackgroundStage && fftSize > m_maxRealtimeFFTSize)
1.123 + fftSize = m_maxRealtimeFFTSize;
1.124 + if (fftSize > m_maxFFTSize)
1.125 + fftSize = m_maxFFTSize;
1.126 + }
1.127 +
1.128 + // Start up background thread
1.129 + // FIXME: would be better to up the thread priority here. It doesn't need to be real-time, but higher than the default...
1.130 + if (this->useBackgroundThreads() && m_backgroundStages.Length() > 0) {
1.131 + if (!m_backgroundThread.Start()) {
1.132 + NS_WARNING("Cannot start convolver thread.");
1.133 + return;
1.134 + }
1.135 + CancelableTask* task = NewRunnableMethod(this, &ReverbConvolver::backgroundThreadEntry);
1.136 + m_backgroundThread.message_loop()->PostTask(FROM_HERE, task);
1.137 + }
1.138 +}
1.139 +
1.140 +ReverbConvolver::~ReverbConvolver()
1.141 +{
1.142 + // Wait for background thread to stop
1.143 + if (useBackgroundThreads() && m_backgroundThread.IsRunning()) {
1.144 + m_wantsToExit = true;
1.145 +
1.146 + // Wake up thread so it can return
1.147 + {
1.148 + AutoLock locker(m_backgroundThreadLock);
1.149 + m_moreInputBuffered = true;
1.150 + m_backgroundThreadCondition.Signal();
1.151 + }
1.152 +
1.153 + m_backgroundThread.Stop();
1.154 + }
1.155 +}
1.156 +
1.157 +size_t ReverbConvolver::sizeOfIncludingThis(mozilla::MallocSizeOf aMallocSizeOf) const
1.158 +{
1.159 + size_t amount = aMallocSizeOf(this);
1.160 + amount += m_stages.SizeOfExcludingThis(aMallocSizeOf);
1.161 + for (size_t i = 0; i < m_stages.Length(); i++) {
1.162 + if (m_stages[i]) {
1.163 + amount += m_stages[i]->sizeOfIncludingThis(aMallocSizeOf);
1.164 + }
1.165 + }
1.166 +
1.167 + amount += m_backgroundStages.SizeOfExcludingThis(aMallocSizeOf);
1.168 + for (size_t i = 0; i < m_backgroundStages.Length(); i++) {
1.169 + if (m_backgroundStages[i]) {
1.170 + amount += m_backgroundStages[i]->sizeOfIncludingThis(aMallocSizeOf);
1.171 + }
1.172 + }
1.173 +
1.174 + // NB: The buffer sizes are static, so even though they might be accessed
1.175 + // in another thread it's safe to measure them.
1.176 + amount += m_accumulationBuffer.sizeOfExcludingThis(aMallocSizeOf);
1.177 + amount += m_inputBuffer.sizeOfExcludingThis(aMallocSizeOf);
1.178 +
1.179 + // Possible future measurements:
1.180 + // - m_backgroundThread
1.181 + // - m_backgroundThreadLock
1.182 + // - m_backgroundThreadCondition
1.183 + return amount;
1.184 +}
1.185 +
1.186 +void ReverbConvolver::backgroundThreadEntry()
1.187 +{
1.188 + while (!m_wantsToExit) {
1.189 + // Wait for realtime thread to give us more input
1.190 + m_moreInputBuffered = false;
1.191 + {
1.192 + AutoLock locker(m_backgroundThreadLock);
1.193 + while (!m_moreInputBuffered && !m_wantsToExit)
1.194 + m_backgroundThreadCondition.Wait();
1.195 + }
1.196 +
1.197 + // Process all of the stages until their read indices reach the input buffer's write index
1.198 + int writeIndex = m_inputBuffer.writeIndex();
1.199 +
1.200 + // Even though it doesn't seem like every stage needs to maintain its own version of readIndex
1.201 + // we do this in case we want to run in more than one background thread.
1.202 + int readIndex;
1.203 +
1.204 + while ((readIndex = m_backgroundStages[0]->inputReadIndex()) != writeIndex) { // FIXME: do better to detect buffer overrun...
1.205 + // The ReverbConvolverStages need to process in amounts which evenly divide half the FFT size
1.206 + const int SliceSize = MinFFTSize / 2;
1.207 +
1.208 + // Accumulate contributions from each stage
1.209 + for (size_t i = 0; i < m_backgroundStages.Length(); ++i)
1.210 + m_backgroundStages[i]->processInBackground(this, SliceSize);
1.211 + }
1.212 + }
1.213 +}
1.214 +
1.215 +void ReverbConvolver::process(const float* sourceChannelData, size_t sourceChannelLength,
1.216 + float* destinationChannelData, size_t destinationChannelLength,
1.217 + size_t framesToProcess)
1.218 +{
1.219 + bool isSafe = sourceChannelData && destinationChannelData && sourceChannelLength >= framesToProcess && destinationChannelLength >= framesToProcess;
1.220 + MOZ_ASSERT(isSafe);
1.221 + if (!isSafe)
1.222 + return;
1.223 +
1.224 + const float* source = sourceChannelData;
1.225 + float* destination = destinationChannelData;
1.226 + bool isDataSafe = source && destination;
1.227 + MOZ_ASSERT(isDataSafe);
1.228 + if (!isDataSafe)
1.229 + return;
1.230 +
1.231 + // Feed input buffer (read by all threads)
1.232 + m_inputBuffer.write(source, framesToProcess);
1.233 +
1.234 + // Accumulate contributions from each stage
1.235 + for (size_t i = 0; i < m_stages.Length(); ++i)
1.236 + m_stages[i]->process(source, framesToProcess);
1.237 +
1.238 + // Finally read from accumulation buffer
1.239 + m_accumulationBuffer.readAndClear(destination, framesToProcess);
1.240 +
1.241 + // Now that we've buffered more input, wake up our background thread.
1.242 +
1.243 + // Not using a MutexLocker looks strange, but we use a tryLock() instead because this is run on the real-time
1.244 + // thread where it is a disaster for the lock to be contended (causes audio glitching). It's OK if we fail to
1.245 + // signal from time to time, since we'll get to it the next time we're called. We're called repeatedly
1.246 + // and frequently (around every 3ms). The background thread is processing well into the future and has a considerable amount of
1.247 + // leeway here...
1.248 + if (m_backgroundThreadLock.Try()) {
1.249 + m_moreInputBuffered = true;
1.250 + m_backgroundThreadCondition.Signal();
1.251 + m_backgroundThreadLock.Release();
1.252 + }
1.253 +}
1.254 +
1.255 +void ReverbConvolver::reset()
1.256 +{
1.257 + for (size_t i = 0; i < m_stages.Length(); ++i)
1.258 + m_stages[i]->reset();
1.259 +
1.260 + for (size_t i = 0; i < m_backgroundStages.Length(); ++i)
1.261 + m_backgroundStages[i]->reset();
1.262 +
1.263 + m_accumulationBuffer.reset();
1.264 + m_inputBuffer.reset();
1.265 +}
1.266 +
1.267 +size_t ReverbConvolver::latencyFrames() const
1.268 +{
1.269 + return 0;
1.270 +}
1.271 +
1.272 +} // namespace WebCore
